Explainable Artificial Intelligence

We will try to explain why the model made a specific prediction on some examples

import pickle
import dalex as dx # XAI library
import matplotlib.pyplot as plt
%matplotlib inline
import scikitplot as skplt
import pandas as pd
import numpy as np
rng = np.random.default_rng(11)

train_data = pd.read_csv('../Data/preprocessed2-train-bank-data.csv', sep=';')
test_data = pd.read_csv('../Data/preprocessed2-test-bank-data.csv', sep=';')

train_data.head()

	age	campaign	contacted.in.previous	emp.var.rate	cons.price.idx	cons.conf.idx	euribor3m	nr.employed	job_admin.	job_blue-collar	...	contact_cellular	contact_telephone	poutcome_failure	poutcome_nonexistent	poutcome_success	month_sin	month_cos	day_sin	day_cos	y
0	0.322632	1.000000	0.0	1.000000	0.612813	0.390735	0.970664	1.000000	0.0	1.0	...	1.0	0.0	0.0	1.0	0.0	-0.500000	-0.866025	-5.877853e-01	-0.809017	0.0
1	0.552137	0.000000	0.0	1.000000	0.422695	0.724448	0.970664	1.000000	0.0	0.0	...	1.0	0.0	0.0	1.0	0.0	-0.866025	-0.500000	-9.510565e-01	0.309017	0.0
2	0.636314	0.495659	0.0	0.854914	0.644951	0.710269	0.932652	0.710922	1.0	0.0	...	0.0	1.0	0.0	1.0	0.0	0.500000	-0.866025	-2.449294e-16	1.000000	0.0
3	0.798640	0.000000	0.0	0.432174	0.331438	0.428275	0.714556	0.743803	0.0	1.0	...	1.0	0.0	1.0	0.0	0.0	-0.500000	0.866025	9.510565e-01	0.309017	0.0
4	0.552137	0.495659	0.0	1.000000	0.612813	0.390735	0.970309	1.000000	0.0	0.0	...	1.0	0.0	0.0	1.0	0.0	-0.500000	-0.866025	-9.510565e-01	0.309017	0.0

5 rows × 44 columns

X_train, y_train = train_data.drop('y', axis=1), train_data['y']
X_test, y_test = test_data.drop('y', axis=1), test_data['y']

"""Loading pickled Random Forest Model"""
file = open('../Models/Serialized_models/random_forest_classifier_gs.pickle', 'rb')
rfc = pickle.load(file)
file.close()

print(rfc)

RandomForestClassifier(class_weight='balanced', criterion='entropy',
                       max_depth=9, min_samples_leaf=3, n_estimators=80,
                       random_state=1)

explainer = dx.Explainer(rfc, X_test, y_test)

Preparation of a new explainer is initiated

  -> data              : 7975 rows 43 cols
  -> target variable   : Parameter 'y' was a pandas.Series. Converted to a numpy.ndarray.
  -> target variable   : 7975 values
  -> model_class       : sklearn.ensemble._forest.RandomForestClassifier (default)
  -> label             : Not specified, model's class short name will be used. (default)
  -> predict function  : <function yhat_proba_default at 0x000001C15158A288> will be used (default)
  -> predict function  : Accepts pandas.DataFrame and numpy.ndarray.
  -> predicted values  : min = 0.135, mean = 0.369, max = 0.977
  -> model type        : classification will be used (default)
  -> residual function : difference between y and yhat (default)
  -> residuals         : min = -0.974, mean = -0.26, max = 0.844
  -> model_info        : package sklearn

A new explainer has been created!

C:\Users\frane\anaconda3\envs\kaggleproject\lib\site-packages\sklearn\base.py:451: UserWarning: X does not have valid feature names, but RandomForestClassifier was fitted with feature names
  "X does not have valid feature names, but"

Example Shap plots

Shap plots can help us understand why model made a specific prediction in the case. It shows which features had a biggest impact on a classification and was the impact positive or negative.

examples = rng.integers(low=0,high=X_test.shape[0]-1,size=4)

for i in examples:
    shap = explainer.predict_parts(X_test.loc[i], type='shap', random_state=1)
    shap.plot(max_vars=30)

'''Predicted and actual target value for plotted examples'''
examples_df = pd.DataFrame([examples,rfc.predict(X_test.iloc[examples]),y_test.iloc[examples]]).transpose()
examples_df.columns = ['index','y_pred','y']
examples_df = examples_df.set_index('index')
examples_df

	y_pred	y
index
1066.0	0.0	0.0
1025.0	1.0	0.0
6355.0	0.0	0.0
3981.0	0.0	0.0

• 1st and 4th examples are pretty similar. In both cases model correctly predicted 0 and economic attributes had the biggest impact on the prediction
• In the 3rd case model also correctly predicted 0, but the main reason for the prediction was type of contact. However, economic attributes also had significant negative impact on the prediction
• In the 2nd case model made a mistake predicting 1. As we can see most features had a positive impact, but despite that customer haven't placed the deposit. There are 2 most likely explanations for that:
  • some factor not included in data influenced client decision
  • this is a result of model sensitivity (earlier we decided to make more sensitive model to improve recall at the expense of precision and accuracy)

Ceteris paribus plots

It shows average probability prediction for individual variable.

pdp = explainer.model_profile(variables=['age','emp.var.rate','cons.price.idx','cons.conf.idx','euribor3m','nr.employed',
                                        'month_sin','month_cos','day_sin','day_cos'], N=X_test.shape[0])
pdp.plot()

Calculating ceteris paribus: 100%|█████████████████████████████████████████████████████| 10/10 [00:37<00:00,  3.71s/it]

Feature importance in the model

permutation_importance = explainer.model_parts(loss_function='rmse', random_state=1)
permutation_importance.plot()

Lift and Cumulative Gains curves

Lift curve shows how many times better it is to use a model than a random choice

skplt.metrics.plot_lift_curve(y_test, rfc.predict_proba(X_test))
plt.legend(loc='upper right')
plt.show()

For example, we see that contacting customers with the top 20% predicted probabilities is more than 3 times better than randomly selecting them.

skplt.metrics.plot_cumulative_gain(y_test, rfc.predict_proba(X_test))
plt.legend(loc='lower right')
plt.show()

In our case it shows how many responders will we reach after contacting some ammount of all of the customers with the highest predicted probabilities.

For example after contacting customers with top 20% predicted probabilities, we will reach over 60% of customers that will place the advertised deposit.